meta-analysis.cz: harmonised estimate-level data from meta-analyses in economics

Estimate-level data from 47 meta-analyses in economics and the social sciences, and a pooled table that stacks 43 of them into one schema. Every value is traceable to the published file it came from, and every published file is the one its authors released. What this release contains. 67,685 rows across the 47 converted source files. The pooled table's primary analysis set is 49,743 estimates satisfying the source papers' own sample definitions. A further 3,136 estimates that those papers excluded are carried alongside for robustness work, flagged rather than hidden, giving 52,879 rows in all. The 52,879 is not "all estimates": it is a selection from the 67,685 source rows. What is new in 2.1.0. One literature is added. No estimate or standard error already in the table changed. social_comparison joins as the 47th dataset and 43rd pooled literature, from Bartos, Irsova, Havranek and Wagenmakers (2025): 79 effects, as Hedges' g, from randomised controlled trials of social comparison as a behaviour change technique, reported in 63 studies. 37 effects are measured against passive and 42 against active control conditions. The paper corrects the two comparisons separately and never pools them, so they are told apart by outcome_variant (vs_passive_control, vs_active_control). All 79 lie inside the paper's own sample. effect_units now names the unit class for five literatures: elasticity for armington and migrant, semi-elasticity for house_prices and price_puzzle, and regression coefficient for habits. armington's effect_alt_units follows. The values themselves are untouched, and migrant's direction_note still records that its stored value is the negative inverse of the elasticity. datasets.json lists chemnitz, a teaching copy of 1,647 short-run electricity estimates, under excluded_resources, so that it is not mistaken for a separate literature. Compatibility with 2.0.0. All 52,800 rows of 2.0.0 are present with the same identifiers, estimates and standard errors; on those rows only the two label columns above differ. A result computed on 2.0.0 reproduces on 2.1.0 once social_comparison is set aside. 2.0.0 was the last breaking release. Anyone moving from 1.3.0 or earlier should read its record, https://doi.org/10.5281/zenodo.22647394. In brief: estimates a paper excluded are carried and flagged by in_paper_sample and paper_sample_exclusion rather than deleted; a second effect scale is carried in effect_alt, se_alt, effect_alt_units and effect_alt_role, and the old pcc and se_pcc columns are retired; class publishes the test-score scale its paper's main models use rather than an appendix robustness metric; twenty activism estimates that carried the wrong sign are corrected; and armington's headline is the winsorised pair. The move from 1.3.0's 50,441 rows to 2.0.0's 52,800 came from these construction rules, not from new evidence. What has not changed. The source files of the 46 earlier datasets, the round-trip guarantee that every harmonised value occurs in its original published file, and the licence. Everything here is CC BY 4.0.